Entropy-Aware Diffusion for Robust Protein Binder Design

Augment RFdiffusion with an entropy-aware loss that penalizes excessive conformational restriction or solvent entropy loss upon binding. Use physics-guided neural networks for entropy prediction and end-point simulations to provide differentiable entropy proxies during training. Couple structure generation (RFdiffusion) with sequence optimization (ProteinMPNN), then validate predicted ΔG decomposition (ΔH, −TΔS) against ITC-calibrated datasets. This approach addresses a major blind spot in current generative design pipelines that focus mainly on enthalpy, by introducing learned, physics-regularized entropy signals into the generative objective. It builds on deep learning binder design successes and extends RFdiffusion by injecting interpretable physics-based constraints, leveraging a hybrid ML + physics strategy to improve generalization and reduce overfitting. Entropy-aware designs are expected to be less brittle to temperature, salt, or minor backbone shifts, enhancing therapeutic robustness and performance on targets with conformational plasticity. The impact is a new class of thermodynamically robust designs with higher experimental success rates, fewer false positives, and better transfer across assay conditions.

References:

1. Calculating the Binding Entropy of Host-Guest Systems with Physics-Guided Neural Networks. Alles Rebel, Ali Risheh, Negin Massoudian, Negin Forouzesh (2022). IEEE International Conference on Bioinformatics and Biomedicine.
2. Machine learning for functional protein design. Pascal Notin, Nathan J. Rollins, Yarin Gal, Chris Sander, Debora Marks (2024). Nature Biotechnology.
3. In silico evolution of protein binders with deep learning models for structure prediction and sequence design. Odessa J Goudy, Amrita Nallathambi, Tomoaki Kinjo, Nicholas Z Randolph, B. Kuhlman (2023). bioRxiv.
4. Broadly applicable and accurate protein design by integrating structure prediction networks and diffusion generative models. Joseph L. Watson, David Juergens, N. Bennett, Brian L. Trippe, Jason Yim, Helen E. Eisenach, Woody Ahern, Andrew J. Borst, R. Ragotte, L. Milles, B. Wicky, Nikita Hanikel, S. Pellock, A. Courbet, W. Sheffler, Jue Wang, Preetham Venkatesh, Isaac Sappington, Susana Vázquez Torres, A. Lauko, Valentin De Bortoli, Emile Mathieu, R. Barzilay, T. Jaakkola, F. DiMaio, M. Baek, D. Baker (2022). bioRxiv.
5. De novo design of protein structure and function with RFdiffusion. Joseph L. Watson, David Juergens, N. Bennett, Brian L. Trippe, Jason Yim, Helen E. Eisenach, Woody Ahern, Andrew J. Borst, Robert J. Ragotte, L. Milles, B. Wicky, Nikita Hanikel, S. Pellock, A. Courbet, W. Sheffler, Jue Wang, Preetham Venkatesh, Isaac Sappington, Susana Vázquez Torres, Anna Lauko, Valentin De Bortoli, Emile Mathieu, Sergey Ovchinnikov, R. Barzilay, T. Jaakkola, Frank Dimaio, M. Baek, David Baker (2023). Nature.